fs/ceph/messenger.c

   1 #include "ceph_debug.h"
   2
   3 #include <linux/crc32c.h>
   4 #include <linux/ctype.h>
   5 #include <linux/highmem.h>
   6 #include <linux/inet.h>
   7 #include <linux/kthread.h>
   8 #include <linux/net.h>
   9 #include <linux/slab.h>
  10 #include <linux/socket.h>
  11 #include <linux/string.h>
  12 #include <linux/bio.h>
  13 #include <linux/blkdev.h>
  14 #include <net/tcp.h>
  15
  16 #include "super.h"
  17 #include "messenger.h"
  18 #include "decode.h"
  19 #include "pagelist.h"
  20
  21 /*
  22  * Ceph uses the messenger to exchange ceph_msg messages with other
  23  * hosts in the system.  The messenger provides ordered and reliable
  24  * delivery.  We tolerate TCP disconnects by reconnecting (with
  25  * exponential backoff) in the case of a fault (disconnection, bad
  26  * crc, protocol error).  Acks allow sent messages to be discarded by
  27  * the sender.
  28  */
  29
  30 /* static tag bytes (protocol control messages) */
  31 static char tag_msg = CEPH_MSGR_TAG_MSG;
  32 static char tag_ack = CEPH_MSGR_TAG_ACK;
  33 static char tag_keepalive = CEPH_MSGR_TAG_KEEPALIVE;
  34
  35 #ifdef CONFIG_LOCKDEP
  36 static struct lock_class_key socket_class;
  37 #endif
  38
  39
  40 static void queue_con(struct ceph_connection *con);
  41 static void con_work(struct work_struct *);
  42 static void ceph_fault(struct ceph_connection *con);
  43
  44 /*
  45  * nicely render a sockaddr as a string.
  46  */
  47 #define MAX_ADDR_STR 20
  48 #define MAX_ADDR_STR_LEN 60
  49 static char addr_str[MAX_ADDR_STR][MAX_ADDR_STR_LEN];
  50 static DEFINE_SPINLOCK(addr_str_lock);
  51 static int last_addr_str;
  52
  53 const char *pr_addr(const struct sockaddr_storage *ss)
  54 {
  55         int i;
  56         char *s;
  57         struct sockaddr_in *in4 = (void *)ss;
  58         struct sockaddr_in6 *in6 = (void *)ss;
  59
  60         spin_lock(&addr_str_lock);
  61         i = last_addr_str++;
  62         if (last_addr_str == MAX_ADDR_STR)
  63                 last_addr_str = 0;
  64         spin_unlock(&addr_str_lock);
  65         s = addr_str[i];
  66
  67         switch (ss->ss_family) {
  68         case AF_INET:
  69                 snprintf(s, MAX_ADDR_STR_LEN, "%pI4:%u", &in4->sin_addr,
  70                          (unsigned int)ntohs(in4->sin_port));
  71                 break;
  72
  73         case AF_INET6:
  74                 snprintf(s, MAX_ADDR_STR_LEN, "[%pI6c]:%u", &in6->sin6_addr,
  75                          (unsigned int)ntohs(in6->sin6_port));
  76                 break;
  77
  78         default:
  79                 sprintf(s, "(unknown sockaddr family %d)", (int)ss->ss_family);
  80         }
  81
  82         return s;
  83 }
  84
  85 static void encode_my_addr(struct ceph_messenger *msgr)
  86 {
  87         memcpy(&msgr->my_enc_addr, &msgr->inst.addr, sizeof(msgr->my_enc_addr));
  88         ceph_encode_addr(&msgr->my_enc_addr);
  89 }
  90
  91 /*
  92  * work queue for all reading and writing to/from the socket.
  93  */
  94 struct workqueue_struct *ceph_msgr_wq;
  95
  96 int __init ceph_msgr_init(void)
  97 {
  98         ceph_msgr_wq = create_workqueue("ceph-msgr");
  99         if (IS_ERR(ceph_msgr_wq)) {
 100                 int ret = PTR_ERR(ceph_msgr_wq);
 101                 pr_err("msgr_init failed to create workqueue: %d\n", ret);
 102                 ceph_msgr_wq = NULL;
 103                 return ret;
 104         }
 105         return 0;
 106 }
 107
 108 void ceph_msgr_exit(void)
 109 {
 110         destroy_workqueue(ceph_msgr_wq);
 111 }
 112
 113 void ceph_msgr_flush(void)
 114 {
 115         flush_workqueue(ceph_msgr_wq);
 116 }
 117
 118
 119 /*
 120  * socket callback functions
 121  */
 122
 123 /* data available on socket, or listen socket received a connect */
 124 static void ceph_data_ready(struct sock *sk, int count_unused)
 125 {
 126         struct ceph_connection *con =
 127                 (struct ceph_connection *)sk->sk_user_data;
 128         if (sk->sk_state != TCP_CLOSE_WAIT) {
 129                 dout("ceph_data_ready on %p state = %lu, queueing work\n",
 130                      con, con->state);
 131                 queue_con(con);
 132         }
 133 }
 134
 135 /* socket has buffer space for writing */
 136 static void ceph_write_space(struct sock *sk)
 137 {
 138         struct ceph_connection *con =
 139                 (struct ceph_connection *)sk->sk_user_data;
 140
 141         /* only queue to workqueue if there is data we want to write. */
 142         if (test_bit(WRITE_PENDING, &con->state)) {
 143                 dout("ceph_write_space %p queueing write work\n", con);
 144                 queue_con(con);
 145         } else {
 146                 dout("ceph_write_space %p nothing to write\n", con);
 147         }
 148
 149         /* since we have our own write_space, clear the SOCK_NOSPACE flag */
 150         clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 151 }
 152
 153 /* socket's state has changed */
 154 static void ceph_state_change(struct sock *sk)
 155 {
 156         struct ceph_connection *con =
 157                 (struct ceph_connection *)sk->sk_user_data;
 158
 159         dout("ceph_state_change %p state = %lu sk_state = %u\n",
 160              con, con->state, sk->sk_state);
 161
 162         if (test_bit(CLOSED, &con->state))
 163                 return;
 164
 165         switch (sk->sk_state) {
 166         case TCP_CLOSE:
 167                 dout("ceph_state_change TCP_CLOSE\n");
 168         case TCP_CLOSE_WAIT:
 169                 dout("ceph_state_change TCP_CLOSE_WAIT\n");
 170                 if (test_and_set_bit(SOCK_CLOSED, &con->state) == 0) {
 171                         if (test_bit(CONNECTING, &con->state))
 172                                 con->error_msg = "connection failed";
 173                         else
 174                                 con->error_msg = "socket closed";
 175                         queue_con(con);
 176                 }
 177                 break;
 178         case TCP_ESTABLISHED:
 179                 dout("ceph_state_change TCP_ESTABLISHED\n");
 180                 queue_con(con);
 181                 break;
 182         }
 183 }
 184
 185 /*
 186  * set up socket callbacks
 187  */
 188 static void set_sock_callbacks(struct socket *sock,
 189                                struct ceph_connection *con)
 190 {
 191         struct sock *sk = sock->sk;
 192         sk->sk_user_data = (void *)con;
 193         sk->sk_data_ready = ceph_data_ready;
 194         sk->sk_write_space = ceph_write_space;
 195         sk->sk_state_change = ceph_state_change;
 196 }
 197
 198
 199 /*
 200  * socket helpers
 201  */
 202
 203 /*
 204  * initiate connection to a remote socket.
 205  */
 206 static struct socket *ceph_tcp_connect(struct ceph_connection *con)
 207 {
 208         struct sockaddr_storage *paddr = &con->peer_addr.in_addr;
 209         struct socket *sock;
 210         int ret;
 211
 212         BUG_ON(con->sock);
 213         ret = sock_create_kern(con->peer_addr.in_addr.ss_family, SOCK_STREAM,
 214                                IPPROTO_TCP, &sock);
 215         if (ret)
 216                 return ERR_PTR(ret);
 217         con->sock = sock;
 218         sock->sk->sk_allocation = GFP_NOFS;
 219
 220 #ifdef CONFIG_LOCKDEP
 221         lockdep_set_class(&sock->sk->sk_lock, &socket_class);
 222 #endif
 223
 224         set_sock_callbacks(sock, con);
 225
 226         dout("connect %s\n", pr_addr(&con->peer_addr.in_addr));
 227
 228         ret = sock->ops->connect(sock, (struct sockaddr *)paddr, sizeof(*paddr),
 229                                  O_NONBLOCK);
 230         if (ret == -EINPROGRESS) {
 231                 dout("connect %s EINPROGRESS sk_state = %u\n",
 232                      pr_addr(&con->peer_addr.in_addr),
 233                      sock->sk->sk_state);
 234                 ret = 0;
 235         }
 236         if (ret < 0) {
 237                 pr_err("connect %s error %d\n",
 238                        pr_addr(&con->peer_addr.in_addr), ret);
 239                 sock_release(sock);
 240                 con->sock = NULL;
 241                 con->error_msg = "connect error";
 242         }
 243
 244         if (ret < 0)
 245                 return ERR_PTR(ret);
 246         return sock;
 247 }
 248
 249 static int ceph_tcp_recvmsg(struct socket *sock, void *buf, size_t len)
 250 {
 251         struct kvec iov = {buf, len};
 252         struct msghdr msg = { .msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL };
 253
 254         return kernel_recvmsg(sock, &msg, &iov, 1, len, msg.msg_flags);
 255 }
 256
 257 /*
 258  * write something.  @more is true if caller will be sending more data
 259  * shortly.
 260  */
 261 static int ceph_tcp_sendmsg(struct socket *sock, struct kvec *iov,
 262                      size_t kvlen, size_t len, int more)
 263 {
 264         struct msghdr msg = { .msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL };
 265
 266         if (more)
 267                 msg.msg_flags |= MSG_MORE;
 268         else
 269                 msg.msg_flags |= MSG_EOR;  /* superfluous, but what the hell */
 270
 271         return kernel_sendmsg(sock, &msg, iov, kvlen, len);
 272 }
 273
 274
 275 /*
 276  * Shutdown/close the socket for the given connection.
 277  */
 278 static int con_close_socket(struct ceph_connection *con)
 279 {
 280         int rc;
 281
 282         dout("con_close_socket on %p sock %p\n", con, con->sock);
 283         if (!con->sock)
 284                 return 0;
 285         set_bit(SOCK_CLOSED, &con->state);
 286         rc = con->sock->ops->shutdown(con->sock, SHUT_RDWR);
 287         sock_release(con->sock);
 288         con->sock = NULL;
 289         clear_bit(SOCK_CLOSED, &con->state);
 290         return rc;
 291 }
 292
 293 /*
 294  * Reset a connection.  Discard all incoming and outgoing messages
 295  * and clear *_seq state.
 296  */
 297 static void ceph_msg_remove(struct ceph_msg *msg)
 298 {
 299         list_del_init(&msg->list_head);
 300         ceph_msg_put(msg);
 301 }
 302 static void ceph_msg_remove_list(struct list_head *head)
 303 {
 304         while (!list_empty(head)) {
 305                 struct ceph_msg *msg = list_first_entry(head, struct ceph_msg,
 306                                                         list_head);
 307                 ceph_msg_remove(msg);
 308         }
 309 }
 310
 311 static void reset_connection(struct ceph_connection *con)
 312 {
 313         /* reset connection, out_queue, msg_ and connect_seq */
 314         /* discard existing out_queue and msg_seq */
 315         ceph_msg_remove_list(&con->out_queue);
 316         ceph_msg_remove_list(&con->out_sent);
 317
 318         if (con->in_msg) {
 319                 ceph_msg_put(con->in_msg);
 320                 con->in_msg = NULL;
 321         }
 322
 323         con->connect_seq = 0;
 324         con->out_seq = 0;
 325         if (con->out_msg) {
 326                 ceph_msg_put(con->out_msg);
 327                 con->out_msg = NULL;
 328         }
 329         con->out_keepalive_pending = false;
 330         con->in_seq = 0;
 331         con->in_seq_acked = 0;
 332 }
 333
 334 /*
 335  * mark a peer down.  drop any open connections.
 336  */
 337 void ceph_con_close(struct ceph_connection *con)
 338 {
 339         dout("con_close %p peer %s\n", con, pr_addr(&con->peer_addr.in_addr));
 340         set_bit(CLOSED, &con->state);  /* in case there's queued work */
 341         clear_bit(STANDBY, &con->state);  /* avoid connect_seq bump */
 342         clear_bit(LOSSYTX, &con->state);  /* so we retry next connect */
 343         clear_bit(KEEPALIVE_PENDING, &con->state);
 344         clear_bit(WRITE_PENDING, &con->state);
 345         mutex_lock(&con->mutex);
 346         reset_connection(con);
 347         con->peer_global_seq = 0;
 348         cancel_delayed_work(&con->work);
 349         mutex_unlock(&con->mutex);
 350         queue_con(con);
 351 }
 352
 353 /*
 354  * Reopen a closed connection, with a new peer address.
 355  */
 356 void ceph_con_open(struct ceph_connection *con, struct ceph_entity_addr *addr)
 357 {
 358         dout("con_open %p %s\n", con, pr_addr(&addr->in_addr));
 359         set_bit(OPENING, &con->state);
 360         clear_bit(CLOSED, &con->state);
 361         memcpy(&con->peer_addr, addr, sizeof(*addr));
 362         con->delay = 0;      /* reset backoff memory */
 363         queue_con(con);
 364 }
 365
 366 /*
 367  * return true if this connection ever successfully opened
 368  */
 369 bool ceph_con_opened(struct ceph_connection *con)
 370 {
 371         return con->connect_seq > 0;
 372 }
 373
 374 /*
 375  * generic get/put
 376  */
 377 struct ceph_connection *ceph_con_get(struct ceph_connection *con)
 378 {
 379         dout("con_get %p nref = %d -> %d\n", con,
 380              atomic_read(&con->nref), atomic_read(&con->nref) + 1);
 381         if (atomic_inc_not_zero(&con->nref))
 382                 return con;
 383         return NULL;
 384 }
 385
 386 void ceph_con_put(struct ceph_connection *con)
 387 {
 388         dout("con_put %p nref = %d -> %d\n", con,
 389              atomic_read(&con->nref), atomic_read(&con->nref) - 1);
 390         BUG_ON(atomic_read(&con->nref) == 0);
 391         if (atomic_dec_and_test(&con->nref)) {
 392                 BUG_ON(con->sock);
 393                 kfree(con);
 394         }
 395 }
 396
 397 /*
 398  * initialize a new connection.
 399  */
 400 void ceph_con_init(struct ceph_messenger *msgr, struct ceph_connection *con)
 401 {
 402         dout("con_init %p\n", con);
 403         memset(con, 0, sizeof(*con));
 404         atomic_set(&con->nref, 1);
 405         con->msgr = msgr;
 406         mutex_init(&con->mutex);
 407         INIT_LIST_HEAD(&con->out_queue);
 408         INIT_LIST_HEAD(&con->out_sent);
 409         INIT_DELAYED_WORK(&con->work, con_work);
 410 }
 411
 412
 413 /*
 414  * We maintain a global counter to order connection attempts.  Get
 415  * a unique seq greater than @gt.
 416  */
 417 static u32 get_global_seq(struct ceph_messenger *msgr, u32 gt)
 418 {
 419         u32 ret;
 420
 421         spin_lock(&msgr->global_seq_lock);
 422         if (msgr->global_seq < gt)
 423                 msgr->global_seq = gt;
 424         ret = ++msgr->global_seq;
 425         spin_unlock(&msgr->global_seq_lock);
 426         return ret;
 427 }
 428
 429
 430 /*
 431  * Prepare footer for currently outgoing message, and finish things
 432  * off.  Assumes out_kvec* are already valid.. we just add on to the end.
 433  */
 434 static void prepare_write_message_footer(struct ceph_connection *con, int v)
 435 {
 436         struct ceph_msg *m = con->out_msg;
 437
 438         dout("prepare_write_message_footer %p\n", con);
 439         con->out_kvec_is_msg = true;
 440         con->out_kvec[v].iov_base = &m->footer;
 441         con->out_kvec[v].iov_len = sizeof(m->footer);
 442         con->out_kvec_bytes += sizeof(m->footer);
 443         con->out_kvec_left++;
 444         con->out_more = m->more_to_follow;
 445         con->out_msg_done = true;
 446 }
 447
 448 /*
 449  * Prepare headers for the next outgoing message.
 450  */
 451 static void prepare_write_message(struct ceph_connection *con)
 452 {
 453         struct ceph_msg *m;
 454         int v = 0;
 455
 456         con->out_kvec_bytes = 0;
 457         con->out_kvec_is_msg = true;
 458         con->out_msg_done = false;
 459
 460         /* Sneak an ack in there first?  If we can get it into the same
 461          * TCP packet that's a good thing. */
 462         if (con->in_seq > con->in_seq_acked) {
 463                 con->in_seq_acked = con->in_seq;
 464                 con->out_kvec[v].iov_base = &tag_ack;
 465                 con->out_kvec[v++].iov_len = 1;
 466                 con->out_temp_ack = cpu_to_le64(con->in_seq_acked);
 467                 con->out_kvec[v].iov_base = &con->out_temp_ack;
 468                 con->out_kvec[v++].iov_len = sizeof(con->out_temp_ack);
 469                 con->out_kvec_bytes = 1 + sizeof(con->out_temp_ack);
 470         }
 471
 472         m = list_first_entry(&con->out_queue,
 473                        struct ceph_msg, list_head);
 474         con->out_msg = m;
 475         if (test_bit(LOSSYTX, &con->state)) {
 476                 list_del_init(&m->list_head);
 477         } else {
 478                 /* put message on sent list */
 479                 ceph_msg_get(m);
 480                 list_move_tail(&m->list_head, &con->out_sent);
 481         }
 482
 483         /*
 484          * only assign outgoing seq # if we haven't sent this message
 485          * yet.  if it is requeued, resend with it's original seq.
 486          */
 487         if (m->needs_out_seq) {
 488                 m->hdr.seq = cpu_to_le64(++con->out_seq);
 489                 m->needs_out_seq = false;
 490         }
 491
 492         dout("prepare_write_message %p seq %lld type %d len %d+%d+%d %d pgs\n",
 493              m, con->out_seq, le16_to_cpu(m->hdr.type),
 494              le32_to_cpu(m->hdr.front_len), le32_to_cpu(m->hdr.middle_len),
 495              le32_to_cpu(m->hdr.data_len),
 496              m->nr_pages);
 497         BUG_ON(le32_to_cpu(m->hdr.front_len) != m->front.iov_len);
 498
 499         /* tag + hdr + front + middle */
 500         con->out_kvec[v].iov_base = &tag_msg;
 501         con->out_kvec[v++].iov_len = 1;
 502         con->out_kvec[v].iov_base = &m->hdr;
 503         con->out_kvec[v++].iov_len = sizeof(m->hdr);
 504         con->out_kvec[v++] = m->front;
 505         if (m->middle)
 506                 con->out_kvec[v++] = m->middle->vec;
 507         con->out_kvec_left = v;
 508         con->out_kvec_bytes += 1 + sizeof(m->hdr) + m->front.iov_len +
 509                 (m->middle ? m->middle->vec.iov_len : 0);
 510         con->out_kvec_cur = con->out_kvec;
 511
 512         /* fill in crc (except data pages), footer */
 513         con->out_msg->hdr.crc =
 514                 cpu_to_le32(crc32c(0, (void *)&m->hdr,
 515                                       sizeof(m->hdr) - sizeof(m->hdr.crc)));
 516         con->out_msg->footer.flags = CEPH_MSG_FOOTER_COMPLETE;
 517         con->out_msg->footer.front_crc =
 518                 cpu_to_le32(crc32c(0, m->front.iov_base, m->front.iov_len));
 519         if (m->middle)
 520                 con->out_msg->footer.middle_crc =
 521                         cpu_to_le32(crc32c(0, m->middle->vec.iov_base,
 522                                            m->middle->vec.iov_len));
 523         else
 524                 con->out_msg->footer.middle_crc = 0;
 525         con->out_msg->footer.data_crc = 0;
 526         dout("prepare_write_message front_crc %u data_crc %u\n",
 527              le32_to_cpu(con->out_msg->footer.front_crc),
 528              le32_to_cpu(con->out_msg->footer.middle_crc));
 529
 530         /* is there a data payload? */
 531         if (le32_to_cpu(m->hdr.data_len) > 0) {
 532                 /* initialize page iterator */
 533                 con->out_msg_pos.page = 0;
 534                 if (m->pages)
 535                         con->out_msg_pos.page_pos =
 536                                 le16_to_cpu(m->hdr.data_off) & ~PAGE_MASK;
 537                 else
 538                         con->out_msg_pos.page_pos = 0;
 539                 con->out_msg_pos.data_pos = 0;
 540                 con->out_msg_pos.did_page_crc = 0;
 541                 con->out_more = 1;  /* data + footer will follow */
 542         } else {
 543                 /* no, queue up footer too and be done */
 544                 prepare_write_message_footer(con, v);
 545         }
 546
 547         set_bit(WRITE_PENDING, &con->state);
 548 }
 549
 550 /*
 551  * Prepare an ack.
 552  */
 553 static void prepare_write_ack(struct ceph_connection *con)
 554 {
 555         dout("prepare_write_ack %p %llu -> %llu\n", con,
 556              con->in_seq_acked, con->in_seq);
 557         con->in_seq_acked = con->in_seq;
 558
 559         con->out_kvec[0].iov_base = &tag_ack;
 560         con->out_kvec[0].iov_len = 1;
 561         con->out_temp_ack = cpu_to_le64(con->in_seq_acked);
 562         con->out_kvec[1].iov_base = &con->out_temp_ack;
 563         con->out_kvec[1].iov_len = sizeof(con->out_temp_ack);
 564         con->out_kvec_left = 2;
 565         con->out_kvec_bytes = 1 + sizeof(con->out_temp_ack);
 566         con->out_kvec_cur = con->out_kvec;
 567         con->out_more = 1;  /* more will follow.. eventually.. */
 568         set_bit(WRITE_PENDING, &con->state);
 569 }
 570
 571 /*
 572  * Prepare to write keepalive byte.
 573  */
 574 static void prepare_write_keepalive(struct ceph_connection *con)
 575 {
 576         dout("prepare_write_keepalive %p\n", con);
 577         con->out_kvec[0].iov_base = &tag_keepalive;
 578         con->out_kvec[0].iov_len = 1;
 579         con->out_kvec_left = 1;
 580         con->out_kvec_bytes = 1;
 581         con->out_kvec_cur = con->out_kvec;
 582         set_bit(WRITE_PENDING, &con->state);
 583 }
 584
 585 /*
 586  * Connection negotiation.
 587  */
 588
 589 static void prepare_connect_authorizer(struct ceph_connection *con)
 590 {
 591         void *auth_buf;
 592         int auth_len = 0;
 593         int auth_protocol = 0;
 594
 595         mutex_unlock(&con->mutex);
 596         if (con->ops->get_authorizer)
 597                 con->ops->get_authorizer(con, &auth_buf, &auth_len,
 598                                          &auth_protocol, &con->auth_reply_buf,
 599                                          &con->auth_reply_buf_len,
 600                                          con->auth_retry);
 601         mutex_lock(&con->mutex);
 602
 603         con->out_connect.authorizer_protocol = cpu_to_le32(auth_protocol);
 604         con->out_connect.authorizer_len = cpu_to_le32(auth_len);
 605
 606         con->out_kvec[con->out_kvec_left].iov_base = auth_buf;
 607         con->out_kvec[con->out_kvec_left].iov_len = auth_len;
 608         con->out_kvec_left++;
 609         con->out_kvec_bytes += auth_len;
 610 }
 611
 612 /*
 613  * We connected to a peer and are saying hello.
 614  */
 615 static void prepare_write_banner(struct ceph_messenger *msgr,
 616                                  struct ceph_connection *con)
 617 {
 618         int len = strlen(CEPH_BANNER);
 619
 620         con->out_kvec[0].iov_base = CEPH_BANNER;
 621         con->out_kvec[0].iov_len = len;
 622         con->out_kvec[1].iov_base = &msgr->my_enc_addr;
 623         con->out_kvec[1].iov_len = sizeof(msgr->my_enc_addr);
 624         con->out_kvec_left = 2;
 625         con->out_kvec_bytes = len + sizeof(msgr->my_enc_addr);
 626         con->out_kvec_cur = con->out_kvec;
 627         con->out_more = 0;
 628         set_bit(WRITE_PENDING, &con->state);
 629 }
 630
 631 static void prepare_write_connect(struct ceph_messenger *msgr,
 632                                   struct ceph_connection *con,
 633                                   int after_banner)
 634 {
 635         unsigned global_seq = get_global_seq(con->msgr, 0);
 636         int proto;
 637
 638         switch (con->peer_name.type) {
 639         case CEPH_ENTITY_TYPE_MON:
 640                 proto = CEPH_MONC_PROTOCOL;
 641                 break;
 642         case CEPH_ENTITY_TYPE_OSD:
 643                 proto = CEPH_OSDC_PROTOCOL;
 644                 break;
 645         case CEPH_ENTITY_TYPE_MDS:
 646                 proto = CEPH_MDSC_PROTOCOL;
 647                 break;
 648         default:
 649                 BUG();
 650         }
 651
 652         dout("prepare_write_connect %p cseq=%d gseq=%d proto=%d\n", con,
 653              con->connect_seq, global_seq, proto);
 654
 655         con->out_connect.features = cpu_to_le64(CEPH_FEATURE_SUPPORTED);
 656         con->out_connect.host_type = cpu_to_le32(CEPH_ENTITY_TYPE_CLIENT);
 657         con->out_connect.connect_seq = cpu_to_le32(con->connect_seq);
 658         con->out_connect.global_seq = cpu_to_le32(global_seq);
 659         con->out_connect.protocol_version = cpu_to_le32(proto);
 660         con->out_connect.flags = 0;
 661
 662         if (!after_banner) {
 663                 con->out_kvec_left = 0;
 664                 con->out_kvec_bytes = 0;
 665         }
 666         con->out_kvec[con->out_kvec_left].iov_base = &con->out_connect;
 667         con->out_kvec[con->out_kvec_left].iov_len = sizeof(con->out_connect);
 668         con->out_kvec_left++;
 669         con->out_kvec_bytes += sizeof(con->out_connect);
 670         con->out_kvec_cur = con->out_kvec;
 671         con->out_more = 0;
 672         set_bit(WRITE_PENDING, &con->state);
 673
 674         prepare_connect_authorizer(con);
 675 }
 676
 677
 678 /*
 679  * write as much of pending kvecs to the socket as we can.
 680  *  1 -> done
 681  *  0 -> socket full, but more to do
 682  * <0 -> error
 683  */
 684 static int write_partial_kvec(struct ceph_connection *con)
 685 {
 686         int ret;
 687
 688         dout("write_partial_kvec %p %d left\n", con, con->out_kvec_bytes);
 689         while (con->out_kvec_bytes > 0) {
 690                 ret = ceph_tcp_sendmsg(con->sock, con->out_kvec_cur,
 691                                        con->out_kvec_left, con->out_kvec_bytes,
 692                                        con->out_more);
 693                 if (ret <= 0)
 694                         goto out;
 695                 con->out_kvec_bytes -= ret;
 696                 if (con->out_kvec_bytes == 0)
 697                         break;            /* done */
 698                 while (ret > 0) {
 699                         if (ret >= con->out_kvec_cur->iov_len) {
 700                                 ret -= con->out_kvec_cur->iov_len;
 701                                 con->out_kvec_cur++;
 702                                 con->out_kvec_left--;
 703                         } else {
 704                                 con->out_kvec_cur->iov_len -= ret;
 705                                 con->out_kvec_cur->iov_base += ret;
 706                                 ret = 0;
 707                                 break;
 708                         }
 709                 }
 710         }
 711         con->out_kvec_left = 0;
 712         con->out_kvec_is_msg = false;
 713         ret = 1;
 714 out:
 715         dout("write_partial_kvec %p %d left in %d kvecs ret = %d\n", con,
 716              con->out_kvec_bytes, con->out_kvec_left, ret);
 717         return ret;  /* done! */
 718 }
 719
 720 #ifdef CONFIG_BLOCK
 721 static void init_bio_iter(struct bio *bio, struct bio **iter, int *seg)
 722 {
 723         if (!bio) {
 724                 *iter = NULL;
 725                 *seg = 0;
 726                 return;
 727         }
 728         *iter = bio;
 729         *seg = bio->bi_idx;
 730 }
 731
 732 static void iter_bio_next(struct bio **bio_iter, int *seg)
 733 {
 734         if (*bio_iter == NULL)
 735                 return;
 736
 737         BUG_ON(*seg >= (*bio_iter)->bi_vcnt);
 738
 739         (*seg)++;
 740         if (*seg == (*bio_iter)->bi_vcnt)
 741                 init_bio_iter((*bio_iter)->bi_next, bio_iter, seg);
 742 }
 743 #endif
 744
 745 /*
 746  * Write as much message data payload as we can.  If we finish, queue
 747  * up the footer.
 748  *  1 -> done, footer is now queued in out_kvec[].
 749  *  0 -> socket full, but more to do
 750  * <0 -> error
 751  */
 752 static int write_partial_msg_pages(struct ceph_connection *con)
 753 {
 754         struct ceph_msg *msg = con->out_msg;
 755         unsigned data_len = le32_to_cpu(msg->hdr.data_len);
 756         size_t len;
 757         int crc = con->msgr->nocrc;
 758         int ret;
 759         int total_max_write;
 760         int in_trail = 0;
 761         size_t trail_len = (msg->trail ? msg->trail->length : 0);
 762
 763         dout("write_partial_msg_pages %p msg %p page %d/%d offset %d\n",
 764              con, con->out_msg, con->out_msg_pos.page, con->out_msg->nr_pages,
 765              con->out_msg_pos.page_pos);
 766
 767 #ifdef CONFIG_BLOCK
 768         if (msg->bio && !msg->bio_iter)
 769                 init_bio_iter(msg->bio, &msg->bio_iter, &msg->bio_seg);
 770 #endif
 771
 772         while (data_len > con->out_msg_pos.data_pos) {
 773                 struct page *page = NULL;
 774                 void *kaddr = NULL;
 775                 int max_write = PAGE_SIZE;
 776                 int page_shift = 0;
 777
 778                 total_max_write = data_len - trail_len -
 779                         con->out_msg_pos.data_pos;
 780
 781                 /*
 782                  * if we are calculating the data crc (the default), we need
 783                  * to map the page.  if our pages[] has been revoked, use the
 784                  * zero page.
 785                  */
 786
 787                 /* have we reached the trail part of the data? */
 788                 if (con->out_msg_pos.data_pos >= data_len - trail_len) {
 789                         in_trail = 1;
 790
 791                         total_max_write = data_len - con->out_msg_pos.data_pos;
 792
 793                         page = list_first_entry(&msg->trail->head,
 794                                                 struct page, lru);
 795                         if (crc)
 796                                 kaddr = kmap(page);
 797                         max_write = PAGE_SIZE;
 798                 } else if (msg->pages) {
 799                         page = msg->pages[con->out_msg_pos.page];
 800                         if (crc)
 801                                 kaddr = kmap(page);
 802                 } else if (msg->pagelist) {
 803                         page = list_first_entry(&msg->pagelist->head,
 804                                                 struct page, lru);
 805                         if (crc)
 806                                 kaddr = kmap(page);
 807 #ifdef CONFIG_BLOCK
 808                 } else if (msg->bio) {
 809                         struct bio_vec *bv;
 810
 811                         bv = bio_iovec_idx(msg->bio_iter, msg->bio_seg);
 812                         page = bv->bv_page;
 813                         page_shift = bv->bv_offset;
 814                         if (crc)
 815                                 kaddr = kmap(page) + page_shift;
 816                         max_write = bv->bv_len;
 817 #endif
 818                 } else {
 819                         page = con->msgr->zero_page;
 820                         if (crc)
 821                                 kaddr = page_address(con->msgr->zero_page);
 822                 }
 823                 len = min_t(int, max_write - con->out_msg_pos.page_pos,
 824                             total_max_write);
 825
 826                 if (crc && !con->out_msg_pos.did_page_crc) {
 827                         void *base = kaddr + con->out_msg_pos.page_pos;
 828                         u32 tmpcrc = le32_to_cpu(con->out_msg->footer.data_crc);
 829
 830                         BUG_ON(kaddr == NULL);
 831                         con->out_msg->footer.data_crc =
 832                                 cpu_to_le32(crc32c(tmpcrc, base, len));
 833                         con->out_msg_pos.did_page_crc = 1;
 834                 }
 835                 ret = kernel_sendpage(con->sock, page,
 836                                       con->out_msg_pos.page_pos + page_shift,
 837                                       len,
 838                                       MSG_DONTWAIT | MSG_NOSIGNAL |
 839                                       MSG_MORE);
 840
 841                 if (crc &&
 842                     (msg->pages || msg->pagelist || msg->bio || in_trail))
 843                         kunmap(page);
 844
 845                 if (ret <= 0)
 846                         goto out;
 847
 848                 con->out_msg_pos.data_pos += ret;
 849                 con->out_msg_pos.page_pos += ret;
 850                 if (ret == len) {
 851                         con->out_msg_pos.page_pos = 0;
 852                         con->out_msg_pos.page++;
 853                         con->out_msg_pos.did_page_crc = 0;
 854                         if (in_trail)
 855                                 list_move_tail(&page->lru,
 856                                                &msg->trail->head);
 857                         else if (msg->pagelist)
 858                                 list_move_tail(&page->lru,
 859                                                &msg->pagelist->head);
 860 #ifdef CONFIG_BLOCK
 861                         else if (msg->bio)
 862                                 iter_bio_next(&msg->bio_iter, &msg->bio_seg);
 863 #endif
 864                 }
 865         }
 866
 867         dout("write_partial_msg_pages %p msg %p done\n", con, msg);
 868
 869         /* prepare and queue up footer, too */
 870         if (!crc)
 871                 con->out_msg->footer.flags |= CEPH_MSG_FOOTER_NOCRC;
 872         con->out_kvec_bytes = 0;
 873         con->out_kvec_left = 0;
 874         con->out_kvec_cur = con->out_kvec;
 875         prepare_write_message_footer(con, 0);
 876         ret = 1;
 877 out:
 878         return ret;
 879 }
 880
 881 /*
 882  * write some zeros
 883  */
 884 static int write_partial_skip(struct ceph_connection *con)
 885 {
 886         int ret;
 887
 888         while (con->out_skip > 0) {
 889                 struct kvec iov = {
 890                         .iov_base = page_address(con->msgr->zero_page),
 891                         .iov_len = min(con->out_skip, (int)PAGE_CACHE_SIZE)
 892                 };
 893
 894                 ret = ceph_tcp_sendmsg(con->sock, &iov, 1, iov.iov_len, 1);
 895                 if (ret <= 0)
 896                         goto out;
 897                 con->out_skip -= ret;
 898         }
 899         ret = 1;
 900 out:
 901         return ret;
 902 }
 903
 904 /*
 905  * Prepare to read connection handshake, or an ack.
 906  */
 907 static void prepare_read_banner(struct ceph_connection *con)
 908 {
 909         dout("prepare_read_banner %p\n", con);
 910         con->in_base_pos = 0;
 911 }
 912
 913 static void prepare_read_connect(struct ceph_connection *con)
 914 {
 915         dout("prepare_read_connect %p\n", con);
 916         con->in_base_pos = 0;
 917 }
 918
 919 static void prepare_read_ack(struct ceph_connection *con)
 920 {
 921         dout("prepare_read_ack %p\n", con);
 922         con->in_base_pos = 0;
 923 }
 924
 925 static void prepare_read_tag(struct ceph_connection *con)
 926 {
 927         dout("prepare_read_tag %p\n", con);
 928         con->in_base_pos = 0;
 929         con->in_tag = CEPH_MSGR_TAG_READY;
 930 }
 931
 932 /*
 933  * Prepare to read a message.
 934  */
 935 static int prepare_read_message(struct ceph_connection *con)
 936 {
 937         dout("prepare_read_message %p\n", con);
 938         BUG_ON(con->in_msg != NULL);
 939         con->in_base_pos = 0;
 940         con->in_front_crc = con->in_middle_crc = con->in_data_crc = 0;
 941         return 0;
 942 }
 943
 944
 945 static int read_partial(struct ceph_connection *con,
 946                         int *to, int size, void *object)
 947 {
 948         *to += size;
 949         while (con->in_base_pos < *to) {
 950                 int left = *to - con->in_base_pos;
 951                 int have = size - left;
 952                 int ret = ceph_tcp_recvmsg(con->sock, object + have, left);
 953                 if (ret <= 0)
 954                         return ret;
 955                 con->in_base_pos += ret;
 956         }
 957         return 1;
 958 }
 959
 960
 961 /*
 962  * Read all or part of the connect-side handshake on a new connection
 963  */
 964 static int read_partial_banner(struct ceph_connection *con)
 965 {
 966         int ret, to = 0;
 967
 968         dout("read_partial_banner %p at %d\n", con, con->in_base_pos);
 969
 970         /* peer's banner */
 971         ret = read_partial(con, &to, strlen(CEPH_BANNER), con->in_banner);
 972         if (ret <= 0)
 973                 goto out;
 974         ret = read_partial(con, &to, sizeof(con->actual_peer_addr),
 975                            &con->actual_peer_addr);
 976         if (ret <= 0)
 977                 goto out;
 978         ret = read_partial(con, &to, sizeof(con->peer_addr_for_me),
 979                            &con->peer_addr_for_me);
 980         if (ret <= 0)
 981                 goto out;
 982 out:
 983         return ret;
 984 }
 985
 986 static int read_partial_connect(struct ceph_connection *con)
 987 {
 988         int ret, to = 0;
 989
 990         dout("read_partial_connect %p at %d\n", con, con->in_base_pos);
 991
 992         ret = read_partial(con, &to, sizeof(con->in_reply), &con->in_reply);
 993         if (ret <= 0)
 994                 goto out;
 995         ret = read_partial(con, &to, le32_to_cpu(con->in_reply.authorizer_len),
 996                            con->auth_reply_buf);
 997         if (ret <= 0)
 998                 goto out;
 999
1000         dout("read_partial_connect %p tag %d, con_seq = %u, g_seq = %u\n",
1001              con, (int)con->in_reply.tag,
1002              le32_to_cpu(con->in_reply.connect_seq),
1003              le32_to_cpu(con->in_reply.global_seq));
1004 out:
1005         return ret;
1006
1007 }
1008
1009 /*
1010  * Verify the hello banner looks okay.
1011  */
1012 static int verify_hello(struct ceph_connection *con)
1013 {
1014         if (memcmp(con->in_banner, CEPH_BANNER, strlen(CEPH_BANNER))) {
1015                 pr_err("connect to %s got bad banner\n",
1016                        pr_addr(&con->peer_addr.in_addr));
1017                 con->error_msg = "protocol error, bad banner";
1018                 return -1;
1019         }
1020         return 0;
1021 }
1022
1023 static bool addr_is_blank(struct sockaddr_storage *ss)
1024 {
1025         switch (ss->ss_family) {
1026         case AF_INET:
1027                 return ((struct sockaddr_in *)ss)->sin_addr.s_addr == 0;
1028         case AF_INET6:
1029                 return
1030                      ((struct sockaddr_in6 *)ss)->sin6_addr.s6_addr32[0] == 0 &&
1031                      ((struct sockaddr_in6 *)ss)->sin6_addr.s6_addr32[1] == 0 &&
1032                      ((struct sockaddr_in6 *)ss)->sin6_addr.s6_addr32[2] == 0 &&
1033                      ((struct sockaddr_in6 *)ss)->sin6_addr.s6_addr32[3] == 0;
1034         }
1035         return false;
1036 }
1037
1038 static int addr_port(struct sockaddr_storage *ss)
1039 {
1040         switch (ss->ss_family) {
1041         case AF_INET:
1042                 return ntohs(((struct sockaddr_in *)ss)->sin_port);
1043         case AF_INET6:
1044                 return ntohs(((struct sockaddr_in6 *)ss)->sin6_port);
1045         }
1046         return 0;
1047 }
1048
1049 static void addr_set_port(struct sockaddr_storage *ss, int p)
1050 {
1051         switch (ss->ss_family) {
1052         case AF_INET:
1053                 ((struct sockaddr_in *)ss)->sin_port = htons(p);
1054         case AF_INET6:
1055                 ((struct sockaddr_in6 *)ss)->sin6_port = htons(p);
1056         }
1057 }
1058
1059 /*
1060  * Parse an ip[:port] list into an addr array.  Use the default
1061  * monitor port if a port isn't specified.
1062  */
1063 int ceph_parse_ips(const char *c, const char *end,
1064                    struct ceph_entity_addr *addr,
1065                    int max_count, int *count)
1066 {
1067         int i;
1068         const char *p = c;
1069
1070         dout("parse_ips on '%.*s'\n", (int)(end-c), c);
1071         for (i = 0; i < max_count; i++) {
1072                 const char *ipend;
1073                 struct sockaddr_storage *ss = &addr[i].in_addr;
1074                 struct sockaddr_in *in4 = (void *)ss;
1075                 struct sockaddr_in6 *in6 = (void *)ss;
1076                 int port;
1077                 char delim = ',';
1078
1079                 if (*p == '[') {
1080                         delim = ']';
1081                         p++;
1082                 }
1083
1084                 memset(ss, 0, sizeof(*ss));
1085                 if (in4_pton(p, end - p, (u8 *)&in4->sin_addr.s_addr,
1086                              delim, &ipend))
1087                         ss->ss_family = AF_INET;
1088                 else if (in6_pton(p, end - p, (u8 *)&in6->sin6_addr.s6_addr,
1089                                   delim, &ipend))
1090                         ss->ss_family = AF_INET6;
1091                 else
1092                         goto bad;
1093                 p = ipend;
1094
1095                 if (delim == ']') {
1096                         if (*p != ']') {
1097                                 dout("missing matching ']'\n");
1098                                 goto bad;
1099                         }
1100                         p++;
1101                 }
1102
1103                 /* port? */
1104                 if (p < end && *p == ':') {
1105                         port = 0;
1106                         p++;
1107                         while (p < end && *p >= '0' && *p <= '9') {
1108                                 port = (port * 10) + (*p - '0');
1109                                 p++;
1110                         }
1111                         if (port > 65535 || port == 0)
1112                                 goto bad;
1113                 } else {
1114                         port = CEPH_MON_PORT;
1115                 }
1116
1117                 addr_set_port(ss, port);
1118
1119                 dout("parse_ips got %s\n", pr_addr(ss));
1120
1121                 if (p == end)
1122                         break;
1123                 if (*p != ',')
1124                         goto bad;
1125                 p++;
1126         }
1127
1128         if (p != end)
1129                 goto bad;
1130
1131         if (count)
1132                 *count = i + 1;
1133         return 0;
1134
1135 bad:
1136         pr_err("parse_ips bad ip '%.*s'\n", (int)(end - c), c);
1137         return -EINVAL;
1138 }
1139
1140 static int process_banner(struct ceph_connection *con)
1141 {
1142         dout("process_banner on %p\n", con);
1143
1144         if (verify_hello(con) < 0)
1145                 return -1;
1146
1147         ceph_decode_addr(&con->actual_peer_addr);
1148         ceph_decode_addr(&con->peer_addr_for_me);
1149
1150         /*
1151          * Make sure the other end is who we wanted.  note that the other
1152          * end may not yet know their ip address, so if it's 0.0.0.0, give
1153          * them the benefit of the doubt.
1154          */
1155         if (memcmp(&con->peer_addr, &con->actual_peer_addr,
1156                    sizeof(con->peer_addr)) != 0 &&
1157             !(addr_is_blank(&con->actual_peer_addr.in_addr) &&
1158               con->actual_peer_addr.nonce == con->peer_addr.nonce)) {
1159                 pr_warning("wrong peer, want %s/%d, got %s/%d\n",
1160                            pr_addr(&con->peer_addr.in_addr),
1161                            (int)le32_to_cpu(con->peer_addr.nonce),
1162                            pr_addr(&con->actual_peer_addr.in_addr),
1163                            (int)le32_to_cpu(con->actual_peer_addr.nonce));
1164                 con->error_msg = "wrong peer at address";
1165                 return -1;
1166         }
1167
1168         /*
1169          * did we learn our address?
1170          */
1171         if (addr_is_blank(&con->msgr->inst.addr.in_addr)) {
1172                 int port = addr_port(&con->msgr->inst.addr.in_addr);
1173
1174                 memcpy(&con->msgr->inst.addr.in_addr,
1175                        &con->peer_addr_for_me.in_addr,
1176                        sizeof(con->peer_addr_for_me.in_addr));
1177                 addr_set_port(&con->msgr->inst.addr.in_addr, port);
1178                 encode_my_addr(con->msgr);
1179                 dout("process_banner learned my addr is %s\n",
1180                      pr_addr(&con->msgr->inst.addr.in_addr));
1181         }
1182
1183         set_bit(NEGOTIATING, &con->state);
1184         prepare_read_connect(con);
1185         return 0;
1186 }
1187
1188 static void fail_protocol(struct ceph_connection *con)
1189 {
1190         reset_connection(con);
1191         set_bit(CLOSED, &con->state);  /* in case there's queued work */
1192
1193         mutex_unlock(&con->mutex);
1194         if (con->ops->bad_proto)
1195                 con->ops->bad_proto(con);
1196         mutex_lock(&con->mutex);
1197 }
1198
1199 static int process_connect(struct ceph_connection *con)
1200 {
1201         u64 sup_feat = CEPH_FEATURE_SUPPORTED;
1202         u64 req_feat = CEPH_FEATURE_REQUIRED;
1203         u64 server_feat = le64_to_cpu(con->in_reply.features);
1204
1205         dout("process_connect on %p tag %d\n", con, (int)con->in_tag);
1206
1207         switch (con->in_reply.tag) {
1208         case CEPH_MSGR_TAG_FEATURES:
1209                 pr_err("%s%lld %s feature set mismatch,"
1210                        " my %llx < server's %llx, missing %llx\n",
1211                        ENTITY_NAME(con->peer_name),
1212                        pr_addr(&con->peer_addr.in_addr),
1213                        sup_feat, server_feat, server_feat & ~sup_feat);
1214                 con->error_msg = "missing required protocol features";
1215                 fail_protocol(con);
1216                 return -1;
1217
1218         case CEPH_MSGR_TAG_BADPROTOVER:
1219                 pr_err("%s%lld %s protocol version mismatch,"
1220                        " my %d != server's %d\n",
1221                        ENTITY_NAME(con->peer_name),
1222                        pr_addr(&con->peer_addr.in_addr),
1223                        le32_to_cpu(con->out_connect.protocol_version),
1224                        le32_to_cpu(con->in_reply.protocol_version));
1225                 con->error_msg = "protocol version mismatch";
1226                 fail_protocol(con);
1227                 return -1;
1228
1229         case CEPH_MSGR_TAG_BADAUTHORIZER:
1230                 con->auth_retry++;
1231                 dout("process_connect %p got BADAUTHORIZER attempt %d\n", con,
1232                      con->auth_retry);
1233                 if (con->auth_retry == 2) {
1234                         con->error_msg = "connect authorization failure";
1235                         reset_connection(con);
1236                         set_bit(CLOSED, &con->state);
1237                         return -1;
1238                 }
1239                 con->auth_retry = 1;
1240                 prepare_write_connect(con->msgr, con, 0);
1241                 prepare_read_connect(con);
1242                 break;
1243
1244         case CEPH_MSGR_TAG_RESETSESSION:
1245                 /*
1246                  * If we connected with a large connect_seq but the peer
1247                  * has no record of a session with us (no connection, or
1248                  * connect_seq == 0), they will send RESETSESION to indicate
1249                  * that they must have reset their session, and may have
1250                  * dropped messages.
1251                  */
1252                 dout("process_connect got RESET peer seq %u\n",
1253                      le32_to_cpu(con->in_connect.connect_seq));
1254                 pr_err("%s%lld %s connection reset\n",
1255                        ENTITY_NAME(con->peer_name),
1256                        pr_addr(&con->peer_addr.in_addr));
1257                 reset_connection(con);
1258                 prepare_write_connect(con->msgr, con, 0);
1259                 prepare_read_connect(con);
1260
1261                 /* Tell ceph about it. */
1262                 mutex_unlock(&con->mutex);
1263                 pr_info("reset on %s%lld\n", ENTITY_NAME(con->peer_name));
1264                 if (con->ops->peer_reset)
1265                         con->ops->peer_reset(con);
1266                 mutex_lock(&con->mutex);
1267                 break;
1268
1269         case CEPH_MSGR_TAG_RETRY_SESSION:
1270                 /*
1271                  * If we sent a smaller connect_seq than the peer has, try
1272                  * again with a larger value.
1273                  */
1274                 dout("process_connect got RETRY my seq = %u, peer_seq = %u\n",
1275                      le32_to_cpu(con->out_connect.connect_seq),
1276                      le32_to_cpu(con->in_connect.connect_seq));
1277                 con->connect_seq = le32_to_cpu(con->in_connect.connect_seq);
1278                 prepare_write_connect(con->msgr, con, 0);
1279                 prepare_read_connect(con);
1280                 break;
1281
1282         case CEPH_MSGR_TAG_RETRY_GLOBAL:
1283                 /*
1284                  * If we sent a smaller global_seq than the peer has, try
1285                  * again with a larger value.
1286                  */
1287                 dout("process_connect got RETRY_GLOBAL my %u peer_gseq %u\n",
1288                      con->peer_global_seq,
1289                      le32_to_cpu(con->in_connect.global_seq));
1290                 get_global_seq(con->msgr,
1291                                le32_to_cpu(con->in_connect.global_seq));
1292                 prepare_write_connect(con->msgr, con, 0);
1293                 prepare_read_connect(con);
1294                 break;
1295
1296         case CEPH_MSGR_TAG_READY:
1297                 if (req_feat & ~server_feat) {
1298                         pr_err("%s%lld %s protocol feature mismatch,"
1299                                " my required %llx > server's %llx, need %llx\n",
1300                                ENTITY_NAME(con->peer_name),
1301                                pr_addr(&con->peer_addr.in_addr),
1302                                req_feat, server_feat, req_feat & ~server_feat);
1303                         con->error_msg = "missing required protocol features";
1304                         fail_protocol(con);
1305                         return -1;
1306                 }
1307                 clear_bit(CONNECTING, &con->state);
1308                 con->peer_global_seq = le32_to_cpu(con->in_reply.global_seq);
1309                 con->connect_seq++;
1310                 con->peer_features = server_feat;
1311                 dout("process_connect got READY gseq %d cseq %d (%d)\n",
1312                      con->peer_global_seq,
1313                      le32_to_cpu(con->in_reply.connect_seq),
1314                      con->connect_seq);
1315                 WARN_ON(con->connect_seq !=
1316                         le32_to_cpu(con->in_reply.connect_seq));
1317
1318                 if (con->in_reply.flags & CEPH_MSG_CONNECT_LOSSY)
1319                         set_bit(LOSSYTX, &con->state);
1320
1321                 prepare_read_tag(con);
1322                 break;
1323
1324         case CEPH_MSGR_TAG_WAIT:
1325                 /*
1326                  * If there is a connection race (we are opening
1327                  * connections to each other), one of us may just have
1328                  * to WAIT.  This shouldn't happen if we are the
1329                  * client.
1330                  */
1331                 pr_err("process_connect peer connecting WAIT\n");
1332
1333         default:
1334                 pr_err("connect protocol error, will retry\n");
1335                 con->error_msg = "protocol error, garbage tag during connect";
1336                 return -1;
1337         }
1338         return 0;
1339 }
1340
1341
1342 /*
1343  * read (part of) an ack
1344  */
1345 static int read_partial_ack(struct ceph_connection *con)
1346 {
1347         int to = 0;
1348
1349         return read_partial(con, &to, sizeof(con->in_temp_ack),
1350                             &con->in_temp_ack);
1351 }
1352
1353
1354 /*
1355  * We can finally discard anything that's been acked.
1356  */
1357 static void process_ack(struct ceph_connection *con)
1358 {
1359         struct ceph_msg *m;
1360         u64 ack = le64_to_cpu(con->in_temp_ack);
1361         u64 seq;
1362
1363         while (!list_empty(&con->out_sent)) {
1364                 m = list_first_entry(&con->out_sent, struct ceph_msg,
1365                                      list_head);
1366                 seq = le64_to_cpu(m->hdr.seq);
1367                 if (seq > ack)
1368                         break;
1369                 dout("got ack for seq %llu type %d at %p\n", seq,
1370                      le16_to_cpu(m->hdr.type), m);
1371                 ceph_msg_remove(m);
1372         }
1373         prepare_read_tag(con);
1374 }
1375
1376
1377
1378
1379 static int read_partial_message_section(struct ceph_connection *con,
1380                                         struct kvec *section,
1381                                         unsigned int sec_len, u32 *crc)
1382 {
1383         int ret, left;
1384
1385         BUG_ON(!section);
1386
1387         while (section->iov_len < sec_len) {
1388                 BUG_ON(section->iov_base == NULL);
1389                 left = sec_len - section->iov_len;
1390                 ret = ceph_tcp_recvmsg(con->sock, (char *)section->iov_base +
1391                                        section->iov_len, left);
1392                 if (ret <= 0)
1393                         return ret;
1394                 section->iov_len += ret;
1395                 if (section->iov_len == sec_len)
1396                         *crc = crc32c(0, section->iov_base,
1397                                       section->iov_len);
1398         }
1399
1400         return 1;
1401 }
1402
1403 static struct ceph_msg *ceph_alloc_msg(struct ceph_connection *con,
1404                                 struct ceph_msg_header *hdr,
1405                                 int *skip);
1406
1407
1408 static int read_partial_message_pages(struct ceph_connection *con,
1409                                       struct page **pages,
1410                                       unsigned data_len, int datacrc)
1411 {
1412         void *p;
1413         int ret;
1414         int left;
1415
1416         left = min((int)(data_len - con->in_msg_pos.data_pos),
1417                    (int)(PAGE_SIZE - con->in_msg_pos.page_pos));
1418         /* (page) data */
1419         BUG_ON(pages == NULL);
1420         p = kmap(pages[con->in_msg_pos.page]);
1421         ret = ceph_tcp_recvmsg(con->sock, p + con->in_msg_pos.page_pos,
1422                                left);
1423         if (ret > 0 && datacrc)
1424                 con->in_data_crc =
1425                         crc32c(con->in_data_crc,
1426                                   p + con->in_msg_pos.page_pos, ret);
1427         kunmap(pages[con->in_msg_pos.page]);
1428         if (ret <= 0)
1429                 return ret;
1430         con->in_msg_pos.data_pos += ret;
1431         con->in_msg_pos.page_pos += ret;
1432         if (con->in_msg_pos.page_pos == PAGE_SIZE) {
1433                 con->in_msg_pos.page_pos = 0;
1434                 con->in_msg_pos.page++;
1435         }
1436
1437         return ret;
1438 }
1439
1440 #ifdef CONFIG_BLOCK
1441 static int read_partial_message_bio(struct ceph_connection *con,
1442                                     struct bio **bio_iter, int *bio_seg,
1443                                     unsigned data_len, int datacrc)
1444 {
1445         struct bio_vec *bv = bio_iovec_idx(*bio_iter, *bio_seg);
1446         void *p;
1447         int ret, left;
1448
1449         if (IS_ERR(bv))
1450                 return PTR_ERR(bv);
1451
1452         left = min((int)(data_len - con->in_msg_pos.data_pos),
1453                    (int)(bv->bv_len - con->in_msg_pos.page_pos));
1454
1455         p = kmap(bv->bv_page) + bv->bv_offset;
1456
1457         ret = ceph_tcp_recvmsg(con->sock, p + con->in_msg_pos.page_pos,
1458                                left);
1459         if (ret > 0 && datacrc)
1460                 con->in_data_crc =
1461                         crc32c(con->in_data_crc,
1462                                   p + con->in_msg_pos.page_pos, ret);
1463         kunmap(bv->bv_page);
1464         if (ret <= 0)
1465                 return ret;
1466         con->in_msg_pos.data_pos += ret;
1467         con->in_msg_pos.page_pos += ret;
1468         if (con->in_msg_pos.page_pos == bv->bv_len) {
1469                 con->in_msg_pos.page_pos = 0;
1470                 iter_bio_next(bio_iter, bio_seg);
1471         }
1472
1473         return ret;
1474 }
1475 #endif
1476
1477 /*
1478  * read (part of) a message.
1479  */
1480 static int read_partial_message(struct ceph_connection *con)
1481 {
1482         struct ceph_msg *m = con->in_msg;
1483         int ret;
1484         int to, left;
1485         unsigned front_len, middle_len, data_len, data_off;
1486         int datacrc = con->msgr->nocrc;
1487         int skip;
1488         u64 seq;
1489
1490         dout("read_partial_message con %p msg %p\n", con, m);
1491
1492         /* header */
1493         while (con->in_base_pos < sizeof(con->in_hdr)) {
1494                 left = sizeof(con->in_hdr) - con->in_base_pos;
1495                 ret = ceph_tcp_recvmsg(con->sock,
1496                                        (char *)&con->in_hdr + con->in_base_pos,
1497                                        left);
1498                 if (ret <= 0)
1499                         return ret;
1500                 con->in_base_pos += ret;
1501                 if (con->in_base_pos == sizeof(con->in_hdr)) {
1502                         u32 crc = crc32c(0, (void *)&con->in_hdr,
1503                                  sizeof(con->in_hdr) - sizeof(con->in_hdr.crc));
1504                         if (crc != le32_to_cpu(con->in_hdr.crc)) {
1505                                 pr_err("read_partial_message bad hdr "
1506                                        " crc %u != expected %u\n",
1507                                        crc, con->in_hdr.crc);
1508                                 return -EBADMSG;
1509                         }
1510                 }
1511         }
1512         front_len = le32_to_cpu(con->in_hdr.front_len);
1513         if (front_len > CEPH_MSG_MAX_FRONT_LEN)
1514                 return -EIO;
1515         middle_len = le32_to_cpu(con->in_hdr.middle_len);
1516         if (middle_len > CEPH_MSG_MAX_DATA_LEN)
1517                 return -EIO;
1518         data_len = le32_to_cpu(con->in_hdr.data_len);
1519         if (data_len > CEPH_MSG_MAX_DATA_LEN)
1520                 return -EIO;
1521         data_off = le16_to_cpu(con->in_hdr.data_off);
1522
1523         /* verify seq# */
1524         seq = le64_to_cpu(con->in_hdr.seq);
1525         if ((s64)seq - (s64)con->in_seq < 1) {
1526                 pr_info("skipping %s%lld %s seq %lld, expected %lld\n",
1527                         ENTITY_NAME(con->peer_name),
1528                         pr_addr(&con->peer_addr.in_addr),
1529                         seq, con->in_seq + 1);
1530                 con->in_base_pos = -front_len - middle_len - data_len -
1531                         sizeof(m->footer);
1532                 con->in_tag = CEPH_MSGR_TAG_READY;
1533                 con->in_seq++;
1534                 return 0;
1535         } else if ((s64)seq - (s64)con->in_seq > 1) {
1536                 pr_err("read_partial_message bad seq %lld expected %lld\n",
1537                        seq, con->in_seq + 1);
1538                 con->error_msg = "bad message sequence # for incoming message";
1539                 return -EBADMSG;
1540         }
1541
1542         /* allocate message? */
1543         if (!con->in_msg) {
1544                 dout("got hdr type %d front %d data %d\n", con->in_hdr.type,
1545                      con->in_hdr.front_len, con->in_hdr.data_len);
1546                 skip = 0;
1547                 con->in_msg = ceph_alloc_msg(con, &con->in_hdr, &skip);
1548                 if (skip) {
1549                         /* skip this message */
1550                         dout("alloc_msg said skip message\n");
1551                         BUG_ON(con->in_msg);
1552                         con->in_base_pos = -front_len - middle_len - data_len -
1553                                 sizeof(m->footer);
1554                         con->in_tag = CEPH_MSGR_TAG_READY;
1555                         con->in_seq++;
1556                         return 0;
1557                 }
1558                 if (!con->in_msg) {
1559                         con->error_msg =
1560                                 "error allocating memory for incoming message";
1561                         return -ENOMEM;
1562                 }
1563                 m = con->in_msg;
1564                 m->front.iov_len = 0;    /* haven't read it yet */
1565                 if (m->middle)
1566                         m->middle->vec.iov_len = 0;
1567
1568                 con->in_msg_pos.page = 0;
1569                 if (m->pages)
1570                         con->in_msg_pos.page_pos = data_off & ~PAGE_MASK;
1571                 else
1572                         con->in_msg_pos.page_pos = 0;
1573                 con->in_msg_pos.data_pos = 0;
1574         }
1575
1576         /* front */
1577         ret = read_partial_message_section(con, &m->front, front_len,
1578                                            &con->in_front_crc);
1579         if (ret <= 0)
1580                 return ret;
1581
1582         /* middle */
1583         if (m->middle) {
1584                 ret = read_partial_message_section(con, &m->middle->vec,
1585                                                    middle_len,
1586                                                    &con->in_middle_crc);
1587                 if (ret <= 0)
1588                         return ret;
1589         }
1590 #ifdef CONFIG_BLOCK
1591         if (m->bio && !m->bio_iter)
1592                 init_bio_iter(m->bio, &m->bio_iter, &m->bio_seg);
1593 #endif
1594
1595         /* (page) data */
1596         while (con->in_msg_pos.data_pos < data_len) {
1597                 if (m->pages) {
1598                         ret = read_partial_message_pages(con, m->pages,
1599                                                  data_len, datacrc);
1600                         if (ret <= 0)
1601                                 return ret;
1602 #ifdef CONFIG_BLOCK
1603                 } else if (m->bio) {
1604
1605                         ret = read_partial_message_bio(con,
1606                                                  &m->bio_iter, &m->bio_seg,
1607                                                  data_len, datacrc);
1608                         if (ret <= 0)
1609                                 return ret;
1610 #endif
1611                 } else {
1612                         BUG_ON(1);
1613                 }
1614         }
1615
1616         /* footer */
1617         to = sizeof(m->hdr) + sizeof(m->footer);
1618         while (con->in_base_pos < to) {
1619                 left = to - con->in_base_pos;
1620                 ret = ceph_tcp_recvmsg(con->sock, (char *)&m->footer +
1621                                        (con->in_base_pos - sizeof(m->hdr)),
1622                                        left);
1623                 if (ret <= 0)
1624                         return ret;
1625                 con->in_base_pos += ret;
1626         }
1627         dout("read_partial_message got msg %p %d (%u) + %d (%u) + %d (%u)\n",
1628              m, front_len, m->footer.front_crc, middle_len,
1629              m->footer.middle_crc, data_len, m->footer.data_crc);
1630
1631         /* crc ok? */
1632         if (con->in_front_crc != le32_to_cpu(m->footer.front_crc)) {
1633                 pr_err("read_partial_message %p front crc %u != exp. %u\n",
1634                        m, con->in_front_crc, m->footer.front_crc);
1635                 return -EBADMSG;
1636         }
1637         if (con->in_middle_crc != le32_to_cpu(m->footer.middle_crc)) {
1638                 pr_err("read_partial_message %p middle crc %u != exp %u\n",
1639                        m, con->in_middle_crc, m->footer.middle_crc);
1640                 return -EBADMSG;
1641         }
1642         if (datacrc &&
1643             (m->footer.flags & CEPH_MSG_FOOTER_NOCRC) == 0 &&
1644             con->in_data_crc != le32_to_cpu(m->footer.data_crc)) {
1645                 pr_err("read_partial_message %p data crc %u != exp. %u\n", m,
1646                        con->in_data_crc, le32_to_cpu(m->footer.data_crc));
1647                 return -EBADMSG;
1648         }
1649
1650         return 1; /* done! */
1651 }
1652
1653 /*
1654  * Process message.  This happens in the worker thread.  The callback should
1655  * be careful not to do anything that waits on other incoming messages or it
1656  * may deadlock.
1657  */
1658 static void process_message(struct ceph_connection *con)
1659 {
1660         struct ceph_msg *msg;
1661
1662         msg = con->in_msg;
1663         con->in_msg = NULL;
1664
1665         /* if first message, set peer_name */
1666         if (con->peer_name.type == 0)
1667                 con->peer_name = msg->hdr.src;
1668
1669         con->in_seq++;
1670         mutex_unlock(&con->mutex);
1671
1672         dout("===== %p %llu from %s%lld %d=%s len %d+%d (%u %u %u) =====\n",
1673              msg, le64_to_cpu(msg->hdr.seq),
1674              ENTITY_NAME(msg->hdr.src),
1675              le16_to_cpu(msg->hdr.type),
1676              ceph_msg_type_name(le16_to_cpu(msg->hdr.type)),
1677              le32_to_cpu(msg->hdr.front_len),
1678              le32_to_cpu(msg->hdr.data_len),
1679              con->in_front_crc, con->in_middle_crc, con->in_data_crc);
1680         con->ops->dispatch(con, msg);
1681
1682         mutex_lock(&con->mutex);
1683         prepare_read_tag(con);
1684 }
1685
1686
1687 /*
1688  * Write something to the socket.  Called in a worker thread when the
1689  * socket appears to be writeable and we have something ready to send.
1690  */
1691 static int try_write(struct ceph_connection *con)
1692 {
1693         struct ceph_messenger *msgr = con->msgr;
1694         int ret = 1;
1695
1696         dout("try_write start %p state %lu nref %d\n", con, con->state,
1697              atomic_read(&con->nref));
1698
1699 more:
1700         dout("try_write out_kvec_bytes %d\n", con->out_kvec_bytes);
1701
1702         /* open the socket first? */
1703         if (con->sock == NULL) {
1704                 /*
1705                  * if we were STANDBY and are reconnecting _this_
1706                  * connection, bump connect_seq now.  Always bump
1707                  * global_seq.
1708                  */
1709                 if (test_and_clear_bit(STANDBY, &con->state))
1710                         con->connect_seq++;
1711
1712                 prepare_write_banner(msgr, con);
1713                 prepare_write_connect(msgr, con, 1);
1714                 prepare_read_banner(con);
1715                 set_bit(CONNECTING, &con->state);
1716                 clear_bit(NEGOTIATING, &con->state);
1717
1718                 BUG_ON(con->in_msg);
1719                 con->in_tag = CEPH_MSGR_TAG_READY;
1720                 dout("try_write initiating connect on %p new state %lu\n",
1721                      con, con->state);
1722                 con->sock = ceph_tcp_connect(con);
1723                 if (IS_ERR(con->sock)) {
1724                         con->sock = NULL;
1725                         con->error_msg = "connect error";
1726                         ret = -1;
1727                         goto out;
1728                 }
1729         }
1730
1731 more_kvec:
1732         /* kvec data queued? */
1733         if (con->out_skip) {
1734                 ret = write_partial_skip(con);
1735                 if (ret <= 0)
1736                         goto done;
1737                 if (ret < 0) {
1738                         dout("try_write write_partial_skip err %d\n", ret);
1739                         goto done;
1740                 }
1741         }
1742         if (con->out_kvec_left) {
1743                 ret = write_partial_kvec(con);
1744                 if (ret <= 0)
1745                         goto done;
1746         }
1747
1748         /* msg pages? */
1749         if (con->out_msg) {
1750                 if (con->out_msg_done) {
1751                         ceph_msg_put(con->out_msg);
1752                         con->out_msg = NULL;   /* we're done with this one */
1753                         goto do_next;
1754                 }
1755
1756                 ret = write_partial_msg_pages(con);
1757                 if (ret == 1)
1758                         goto more_kvec;  /* we need to send the footer, too! */
1759                 if (ret == 0)
1760                         goto done;
1761                 if (ret < 0) {
1762                         dout("try_write write_partial_msg_pages err %d\n",
1763                              ret);
1764                         goto done;
1765                 }
1766         }
1767
1768 do_next:
1769         if (!test_bit(CONNECTING, &con->state)) {
1770                 /* is anything else pending? */
1771                 if (!list_empty(&con->out_queue)) {
1772                         prepare_write_message(con);
1773                         goto more;
1774                 }
1775                 if (con->in_seq > con->in_seq_acked) {
1776                         prepare_write_ack(con);
1777                         goto more;
1778                 }
1779                 if (test_and_clear_bit(KEEPALIVE_PENDING, &con->state)) {
1780                         prepare_write_keepalive(con);
1781                         goto more;
1782                 }
1783         }
1784
1785         /* Nothing to do! */
1786         clear_bit(WRITE_PENDING, &con->state);
1787         dout("try_write nothing else to write.\n");
1788 done:
1789         ret = 0;
1790 out:
1791         dout("try_write done on %p\n", con);
1792         return ret;
1793 }
1794
1795
1796
1797 /*
1798  * Read what we can from the socket.
1799  */
1800 static int try_read(struct ceph_connection *con)
1801 {
1802         int ret = -1;
1803
1804         if (!con->sock)
1805                 return 0;
1806
1807         if (test_bit(STANDBY, &con->state))
1808                 return 0;
1809
1810         dout("try_read start on %p\n", con);
1811
1812 more:
1813         dout("try_read tag %d in_base_pos %d\n", (int)con->in_tag,
1814              con->in_base_pos);
1815         if (test_bit(CONNECTING, &con->state)) {
1816                 if (!test_bit(NEGOTIATING, &con->state)) {
1817                         dout("try_read connecting\n");
1818                         ret = read_partial_banner(con);
1819                         if (ret <= 0)
1820                                 goto done;
1821                         if (process_banner(con) < 0) {
1822                                 ret = -1;
1823                                 goto out;
1824                         }
1825                 }
1826                 ret = read_partial_connect(con);
1827                 if (ret <= 0)
1828                         goto done;
1829                 if (process_connect(con) < 0) {
1830                         ret = -1;
1831                         goto out;
1832                 }
1833                 goto more;
1834         }
1835
1836         if (con->in_base_pos < 0) {
1837                 /*
1838                  * skipping + discarding content.
1839                  *
1840                  * FIXME: there must be a better way to do this!
1841                  */
1842                 static char buf[1024];
1843                 int skip = min(1024, -con->in_base_pos);
1844                 dout("skipping %d / %d bytes\n", skip, -con->in_base_pos);
1845                 ret = ceph_tcp_recvmsg(con->sock, buf, skip);
1846                 if (ret <= 0)
1847                         goto done;
1848                 con->in_base_pos += ret;
1849                 if (con->in_base_pos)
1850                         goto more;
1851         }
1852         if (con->in_tag == CEPH_MSGR_TAG_READY) {
1853                 /*
1854                  * what's next?
1855                  */
1856                 ret = ceph_tcp_recvmsg(con->sock, &con->in_tag, 1);
1857                 if (ret <= 0)
1858                         goto done;
1859                 dout("try_read got tag %d\n", (int)con->in_tag);
1860                 switch (con->in_tag) {
1861                 case CEPH_MSGR_TAG_MSG:
1862                         prepare_read_message(con);
1863                         break;
1864                 case CEPH_MSGR_TAG_ACK:
1865                         prepare_read_ack(con);
1866                         break;
1867                 case CEPH_MSGR_TAG_CLOSE:
1868                         set_bit(CLOSED, &con->state);   /* fixme */
1869                         goto done;
1870                 default:
1871                         goto bad_tag;
1872                 }
1873         }
1874         if (con->in_tag == CEPH_MSGR_TAG_MSG) {
1875                 ret = read_partial_message(con);
1876                 if (ret <= 0) {
1877                         switch (ret) {
1878                         case -EBADMSG:
1879                                 con->error_msg = "bad crc";
1880                                 ret = -EIO;
1881                                 goto out;
1882                         case -EIO:
1883                                 con->error_msg = "io error";
1884                                 goto out;
1885                         default:
1886                                 goto done;
1887                         }
1888                 }
1889                 if (con->in_tag == CEPH_MSGR_TAG_READY)
1890                         goto more;
1891                 process_message(con);
1892                 goto more;
1893         }
1894         if (con->in_tag == CEPH_MSGR_TAG_ACK) {
1895                 ret = read_partial_ack(con);
1896                 if (ret <= 0)
1897                         goto done;
1898                 process_ack(con);
1899                 goto more;
1900         }
1901
1902 done:
1903         ret = 0;
1904 out:
1905         dout("try_read done on %p\n", con);
1906         return ret;
1907
1908 bad_tag:
1909         pr_err("try_read bad con->in_tag = %d\n", (int)con->in_tag);
1910         con->error_msg = "protocol error, garbage tag";
1911         ret = -1;
1912         goto out;
1913 }
1914
1915
1916 /*
1917  * Atomically queue work on a connection.  Bump @con reference to
1918  * avoid races with connection teardown.
1919  *
1920  * There is some trickery going on with QUEUED and BUSY because we
1921  * only want a _single_ thread operating on each connection at any
1922  * point in time, but we want to use all available CPUs.
1923  *
1924  * The worker thread only proceeds if it can atomically set BUSY.  It
1925  * clears QUEUED and does it's thing.  When it thinks it's done, it
1926  * clears BUSY, then rechecks QUEUED.. if it's set again, it loops
1927  * (tries again to set BUSY).
1928  *
1929  * To queue work, we first set QUEUED, _then_ if BUSY isn't set, we
1930  * try to queue work.  If that fails (work is already queued, or BUSY)
1931  * we give up (work also already being done or is queued) but leave QUEUED
1932  * set so that the worker thread will loop if necessary.
1933  */
1934 static void queue_con(struct ceph_connection *con)
1935 {
1936         if (test_bit(DEAD, &con->state)) {
1937                 dout("queue_con %p ignoring: DEAD\n",
1938                      con);
1939                 return;
1940         }
1941
1942         if (!con->ops->get(con)) {
1943                 dout("queue_con %p ref count 0\n", con);
1944                 return;
1945         }
1946
1947         set_bit(QUEUED, &con->state);
1948         if (test_bit(BUSY, &con->state)) {
1949                 dout("queue_con %p - already BUSY\n", con);
1950                 con->ops->put(con);
1951         } else if (!queue_work(ceph_msgr_wq, &con->work.work)) {
1952                 dout("queue_con %p - already queued\n", con);
1953                 con->ops->put(con);
1954         } else {
1955                 dout("queue_con %p\n", con);
1956         }
1957 }
1958
1959 /*
1960  * Do some work on a connection.  Drop a connection ref when we're done.
1961  */
1962 static void con_work(struct work_struct *work)
1963 {
1964         struct ceph_connection *con = container_of(work, struct ceph_connection,
1965                                                    work.work);
1966         int backoff = 0;
1967
1968 more:
1969         if (test_and_set_bit(BUSY, &con->state) != 0) {
1970                 dout("con_work %p BUSY already set\n", con);
1971                 goto out;
1972         }
1973         dout("con_work %p start, clearing QUEUED\n", con);
1974         clear_bit(QUEUED, &con->state);
1975
1976         mutex_lock(&con->mutex);
1977
1978         if (test_bit(CLOSED, &con->state)) { /* e.g. if we are replaced */
1979                 dout("con_work CLOSED\n");
1980                 con_close_socket(con);
1981                 goto done;
1982         }
1983         if (test_and_clear_bit(OPENING, &con->state)) {
1984                 /* reopen w/ new peer */
1985                 dout("con_work OPENING\n");
1986                 con_close_socket(con);
1987         }
1988
1989         if (test_and_clear_bit(SOCK_CLOSED, &con->state) ||
1990             try_read(con) < 0 ||
1991             try_write(con) < 0) {
1992                 mutex_unlock(&con->mutex);
1993                 backoff = 1;
1994                 ceph_fault(con);     /* error/fault path */
1995                 goto done_unlocked;
1996         }
1997
1998 done:
1999         mutex_unlock(&con->mutex);
2000
2001 done_unlocked:
2002         clear_bit(BUSY, &con->state);
2003         dout("con->state=%lu\n", con->state);
2004         if (test_bit(QUEUED, &con->state)) {
2005                 if (!backoff || test_bit(OPENING, &con->state)) {
2006                         dout("con_work %p QUEUED reset, looping\n", con);
2007                         goto more;
2008                 }
2009                 dout("con_work %p QUEUED reset, but just faulted\n", con);
2010                 clear_bit(QUEUED, &con->state);
2011         }
2012         dout("con_work %p done\n", con);
2013
2014 out:
2015         con->ops->put(con);
2016 }
2017
2018
2019 /*
2020  * Generic error/fault handler.  A retry mechanism is used with
2021  * exponential backoff
2022  */
2023 static void ceph_fault(struct ceph_connection *con)
2024 {
2025         pr_err("%s%lld %s %s\n", ENTITY_NAME(con->peer_name),
2026                pr_addr(&con->peer_addr.in_addr), con->error_msg);
2027         dout("fault %p state %lu to peer %s\n",
2028              con, con->state, pr_addr(&con->peer_addr.in_addr));
2029
2030         if (test_bit(LOSSYTX, &con->state)) {
2031                 dout("fault on LOSSYTX channel\n");
2032                 goto out;
2033         }
2034
2035         mutex_lock(&con->mutex);
2036         if (test_bit(CLOSED, &con->state))
2037                 goto out_unlock;
2038
2039         con_close_socket(con);
2040
2041         if (con->in_msg) {
2042                 ceph_msg_put(con->in_msg);
2043                 con->in_msg = NULL;
2044         }
2045
2046         /* Requeue anything that hasn't been acked */
2047         list_splice_init(&con->out_sent, &con->out_queue);
2048
2049         /* If there are no messages in the queue, place the connection
2050          * in a STANDBY state (i.e., don't try to reconnect just yet). */
2051         if (list_empty(&con->out_queue) && !con->out_keepalive_pending) {
2052                 dout("fault setting STANDBY\n");
2053                 set_bit(STANDBY, &con->state);
2054         } else {
2055                 /* retry after a delay. */
2056                 if (con->delay == 0)
2057                         con->delay = BASE_DELAY_INTERVAL;
2058                 else if (con->delay < MAX_DELAY_INTERVAL)
2059                         con->delay *= 2;
2060                 dout("fault queueing %p delay %lu\n", con, con->delay);
2061                 con->ops->get(con);
2062                 if (queue_delayed_work(ceph_msgr_wq, &con->work,
2063                                        round_jiffies_relative(con->delay)) == 0)
2064                         con->ops->put(con);
2065         }
2066
2067 out_unlock:
2068         mutex_unlock(&con->mutex);
2069 out:
2070         /*
2071          * in case we faulted due to authentication, invalidate our
2072          * current tickets so that we can get new ones.
2073          */
2074         if (con->auth_retry && con->ops->invalidate_authorizer) {
2075                 dout("calling invalidate_authorizer()\n");
2076                 con->ops->invalidate_authorizer(con);
2077         }
2078
2079         if (con->ops->fault)
2080                 con->ops->fault(con);
2081 }
2082
2083
2084
2085 /*
2086  * create a new messenger instance
2087  */
2088 struct ceph_messenger *ceph_messenger_create(struct ceph_entity_addr *myaddr)
2089 {
2090         struct ceph_messenger *msgr;
2091
2092         msgr = kzalloc(sizeof(*msgr), GFP_KERNEL);
2093         if (msgr == NULL)
2094                 return ERR_PTR(-ENOMEM);
2095
2096         spin_lock_init(&msgr->global_seq_lock);
2097
2098         /* the zero page is needed if a request is "canceled" while the message
2099          * is being written over the socket */
2100         msgr->zero_page = __page_cache_alloc(GFP_KERNEL | __GFP_ZERO);
2101         if (!msgr->zero_page) {
2102                 kfree(msgr);
2103                 return ERR_PTR(-ENOMEM);
2104         }
2105         kmap(msgr->zero_page);
2106
2107         if (myaddr)
2108                 msgr->inst.addr = *myaddr;
2109
2110         /* select a random nonce */
2111         msgr->inst.addr.type = 0;
2112         get_random_bytes(&msgr->inst.addr.nonce, sizeof(msgr->inst.addr.nonce));
2113         encode_my_addr(msgr);
2114
2115         dout("messenger_create %p\n", msgr);
2116         return msgr;
2117 }
2118
2119 void ceph_messenger_destroy(struct ceph_messenger *msgr)
2120 {
2121         dout("destroy %p\n", msgr);
2122         kunmap(msgr->zero_page);
2123         __free_page(msgr->zero_page);
2124         kfree(msgr);
2125         dout("destroyed messenger %p\n", msgr);
2126 }
2127
2128 /*
2129  * Queue up an outgoing message on the given connection.
2130  */
2131 void ceph_con_send(struct ceph_connection *con, struct ceph_msg *msg)
2132 {
2133         if (test_bit(CLOSED, &con->state)) {
2134                 dout("con_send %p closed, dropping %p\n", con, msg);
2135                 ceph_msg_put(msg);
2136                 return;
2137         }
2138
2139         /* set src+dst */
2140         msg->hdr.src = con->msgr->inst.name;
2141
2142         BUG_ON(msg->front.iov_len != le32_to_cpu(msg->hdr.front_len));
2143
2144         msg->needs_out_seq = true;
2145
2146         /* queue */
2147         mutex_lock(&con->mutex);
2148         BUG_ON(!list_empty(&msg->list_head));
2149         list_add_tail(&msg->list_head, &con->out_queue);
2150         dout("----- %p to %s%lld %d=%s len %d+%d+%d -----\n", msg,
2151              ENTITY_NAME(con->peer_name), le16_to_cpu(msg->hdr.type),
2152              ceph_msg_type_name(le16_to_cpu(msg->hdr.type)),
2153              le32_to_cpu(msg->hdr.front_len),
2154              le32_to_cpu(msg->hdr.middle_len),
2155              le32_to_cpu(msg->hdr.data_len));
2156         mutex_unlock(&con->mutex);
2157
2158         /* if there wasn't anything waiting to send before, queue
2159          * new work */
2160         if (test_and_set_bit(WRITE_PENDING, &con->state) == 0)
2161                 queue_con(con);
2162 }
2163
2164 /*
2165  * Revoke a message that was previously queued for send
2166  */
2167 void ceph_con_revoke(struct ceph_connection *con, struct ceph_msg *msg)
2168 {
2169         mutex_lock(&con->mutex);
2170         if (!list_empty(&msg->list_head)) {
2171                 dout("con_revoke %p msg %p - was on queue\n", con, msg);
2172                 list_del_init(&msg->list_head);
2173                 ceph_msg_put(msg);
2174                 msg->hdr.seq = 0;
2175         }
2176         if (con->out_msg == msg) {
2177                 dout("con_revoke %p msg %p - was sending\n", con, msg);
2178                 con->out_msg = NULL;
2179                 if (con->out_kvec_is_msg) {
2180                         con->out_skip = con->out_kvec_bytes;
2181                         con->out_kvec_is_msg = false;
2182                 }
2183                 ceph_msg_put(msg);
2184                 msg->hdr.seq = 0;
2185         }
2186         mutex_unlock(&con->mutex);
2187 }
2188
2189 /*
2190  * Revoke a message that we may be reading data into
2191  */
2192 void ceph_con_revoke_message(struct ceph_connection *con, struct ceph_msg *msg)
2193 {
2194         mutex_lock(&con->mutex);
2195         if (con->in_msg && con->in_msg == msg) {
2196                 unsigned front_len = le32_to_cpu(con->in_hdr.front_len);
2197                 unsigned middle_len = le32_to_cpu(con->in_hdr.middle_len);
2198                 unsigned data_len = le32_to_cpu(con->in_hdr.data_len);
2199
2200                 /* skip rest of message */
2201                 dout("con_revoke_pages %p msg %p revoked\n", con, msg);
2202                         con->in_base_pos = con->in_base_pos -
2203                                 sizeof(struct ceph_msg_header) -
2204                                 front_len -
2205                                 middle_len -
2206                                 data_len -
2207                                 sizeof(struct ceph_msg_footer);
2208                 ceph_msg_put(con->in_msg);
2209                 con->in_msg = NULL;
2210                 con->in_tag = CEPH_MSGR_TAG_READY;
2211                 con->in_seq++;
2212         } else {
2213                 dout("con_revoke_pages %p msg %p pages %p no-op\n",
2214                      con, con->in_msg, msg);
2215         }
2216         mutex_unlock(&con->mutex);
2217 }
2218
2219 /*
2220  * Queue a keepalive byte to ensure the tcp connection is alive.
2221  */
2222 void ceph_con_keepalive(struct ceph_connection *con)
2223 {
2224         if (test_and_set_bit(KEEPALIVE_PENDING, &con->state) == 0 &&
2225             test_and_set_bit(WRITE_PENDING, &con->state) == 0)
2226                 queue_con(con);
2227 }
2228
2229
2230 /*
2231  * construct a new message with given type, size
2232  * the new msg has a ref count of 1.
2233  */
2234 struct ceph_msg *ceph_msg_new(int type, int front_len, gfp_t flags)
2235 {
2236         struct ceph_msg *m;
2237
2238         m = kmalloc(sizeof(*m), flags);
2239         if (m == NULL)
2240                 goto out;
2241         kref_init(&m->kref);
2242         INIT_LIST_HEAD(&m->list_head);
2243
2244         m->hdr.tid = 0;
2245         m->hdr.type = cpu_to_le16(type);
2246         m->hdr.priority = cpu_to_le16(CEPH_MSG_PRIO_DEFAULT);
2247         m->hdr.version = 0;
2248         m->hdr.front_len = cpu_to_le32(front_len);
2249         m->hdr.middle_len = 0;
2250         m->hdr.data_len = 0;
2251         m->hdr.data_off = 0;
2252         m->hdr.reserved = 0;
2253         m->footer.front_crc = 0;
2254         m->footer.middle_crc = 0;
2255         m->footer.data_crc = 0;
2256         m->footer.flags = 0;
2257         m->front_max = front_len;
2258         m->front_is_vmalloc = false;
2259         m->more_to_follow = false;
2260         m->pool = NULL;
2261
2262         /* front */
2263         if (front_len) {
2264                 if (front_len > PAGE_CACHE_SIZE) {
2265                         m->front.iov_base = __vmalloc(front_len, flags,
2266                                                       PAGE_KERNEL);
2267                         m->front_is_vmalloc = true;
2268                 } else {
2269                         m->front.iov_base = kmalloc(front_len, flags);
2270                 }
2271                 if (m->front.iov_base == NULL) {
2272                         pr_err("msg_new can't allocate %d bytes\n",
2273                              front_len);
2274                         goto out2;
2275                 }
2276         } else {
2277                 m->front.iov_base = NULL;
2278         }
2279         m->front.iov_len = front_len;
2280
2281         /* middle */
2282         m->middle = NULL;
2283
2284         /* data */
2285         m->nr_pages = 0;
2286         m->pages = NULL;
2287         m->pagelist = NULL;
2288         m->bio = NULL;
2289         m->bio_iter = NULL;
2290         m->bio_seg = 0;
2291         m->trail = NULL;
2292
2293         dout("ceph_msg_new %p front %d\n", m, front_len);
2294         return m;
2295
2296 out2:
2297         ceph_msg_put(m);
2298 out:
2299         pr_err("msg_new can't create type %d front %d\n", type, front_len);
2300         return NULL;
2301 }
2302
2303 /*
2304  * Allocate "middle" portion of a message, if it is needed and wasn't
2305  * allocated by alloc_msg.  This allows us to read a small fixed-size
2306  * per-type header in the front and then gracefully fail (i.e.,
2307  * propagate the error to the caller based on info in the front) when
2308  * the middle is too large.
2309  */
2310 static int ceph_alloc_middle(struct ceph_connection *con, struct ceph_msg *msg)
2311 {
2312         int type = le16_to_cpu(msg->hdr.type);
2313         int middle_len = le32_to_cpu(msg->hdr.middle_len);
2314
2315         dout("alloc_middle %p type %d %s middle_len %d\n", msg, type,
2316              ceph_msg_type_name(type), middle_len);
2317         BUG_ON(!middle_len);
2318         BUG_ON(msg->middle);
2319
2320         msg->middle = ceph_buffer_new(middle_len, GFP_NOFS);
2321         if (!msg->middle)
2322                 return -ENOMEM;
2323         return 0;
2324 }
2325
2326 /*
2327  * Generic message allocator, for incoming messages.
2328  */
2329 static struct ceph_msg *ceph_alloc_msg(struct ceph_connection *con,
2330                                 struct ceph_msg_header *hdr,
2331                                 int *skip)
2332 {
2333         int type = le16_to_cpu(hdr->type);
2334         int front_len = le32_to_cpu(hdr->front_len);
2335         int middle_len = le32_to_cpu(hdr->middle_len);
2336         struct ceph_msg *msg = NULL;
2337         int ret;
2338
2339         if (con->ops->alloc_msg) {
2340                 mutex_unlock(&con->mutex);
2341                 msg = con->ops->alloc_msg(con, hdr, skip);
2342                 mutex_lock(&con->mutex);
2343                 if (!msg || *skip)
2344                         return NULL;
2345         }
2346         if (!msg) {
2347                 *skip = 0;
2348                 msg = ceph_msg_new(type, front_len, GFP_NOFS);
2349                 if (!msg) {
2350                         pr_err("unable to allocate msg type %d len %d\n",
2351                                type, front_len);
2352                         return NULL;
2353                 }
2354         }
2355         memcpy(&msg->hdr, &con->in_hdr, sizeof(con->in_hdr));
2356
2357         if (middle_len && !msg->middle) {
2358                 ret = ceph_alloc_middle(con, msg);
2359                 if (ret < 0) {
2360                         ceph_msg_put(msg);
2361                         return NULL;
2362                 }
2363         }
2364
2365         return msg;
2366 }
2367
2368
2369 /*
2370  * Free a generically kmalloc'd message.
2371  */
2372 void ceph_msg_kfree(struct ceph_msg *m)
2373 {
2374         dout("msg_kfree %p\n", m);
2375         if (m->front_is_vmalloc)
2376                 vfree(m->front.iov_base);
2377         else
2378                 kfree(m->front.iov_base);
2379         kfree(m);
2380 }
2381
2382 /*
2383  * Drop a msg ref.  Destroy as needed.
2384  */
2385 void ceph_msg_last_put(struct kref *kref)
2386 {
2387         struct ceph_msg *m = container_of(kref, struct ceph_msg, kref);
2388
2389         dout("ceph_msg_put last one on %p\n", m);
2390         WARN_ON(!list_empty(&m->list_head));
2391
2392         /* drop middle, data, if any */
2393         if (m->middle) {
2394                 ceph_buffer_put(m->middle);
2395                 m->middle = NULL;
2396         }
2397         m->nr_pages = 0;
2398         m->pages = NULL;
2399
2400         if (m->pagelist) {
2401                 ceph_pagelist_release(m->pagelist);
2402                 kfree(m->pagelist);
2403                 m->pagelist = NULL;
2404         }
2405
2406         m->trail = NULL;
2407
2408         if (m->pool)
2409                 ceph_msgpool_put(m->pool, m);
2410         else
2411                 ceph_msg_kfree(m);
2412 }
2413
2414 void ceph_msg_dump(struct ceph_msg *msg)
2415 {
2416         pr_debug("msg_dump %p (front_max %d nr_pages %d)\n", msg,
2417                  msg->front_max, msg->nr_pages);
2418         print_hex_dump(KERN_DEBUG, "header: ",
2419                        DUMP_PREFIX_OFFSET, 16, 1,
2420                        &msg->hdr, sizeof(msg->hdr), true);
2421         print_hex_dump(KERN_DEBUG, " front: ",
2422                        DUMP_PREFIX_OFFSET, 16, 1,
2423                        msg->front.iov_base, msg->front.iov_len, true);
2424         if (msg->middle)
2425                 print_hex_dump(KERN_DEBUG, "middle: ",
2426                                DUMP_PREFIX_OFFSET, 16, 1,
2427                                msg->middle->vec.iov_base,
2428                                msg->middle->vec.iov_len, true);
2429         print_hex_dump(KERN_DEBUG, "footer: ",
2430                        DUMP_PREFIX_OFFSET, 16, 1,
2431                        &msg->footer, sizeof(msg->footer), true);
2432 }